Causal Inference
A Novel Two-Step Framework for Estimating Dynamic Treatment Effects Using Longitudinal Covariate Information and Outcome Follow-up Io Murayama* Io Murayama Murayama Murayama Murayama Murayama Murayama Faculty of Medicine, The University of Tokyo
Routinely collected healthcare data often present heterogeneous measurement granularity over follow-up: a short period of intensive longitudinal observation with frequent measurements of treatments, outcomes, and covariates, followed by longer follow-up in which only outcomes can be ascertained. In our intensive care unit (ICU) electronic health record database, high-frequency patient status and treatment information is recorded during the ICU stay, but only mortality is observed after ICU discharge. In such settings, standard causal approaches for time-varying treatments restrict analysis to the period of intensive observation, leaving subsequent outcome information underutilized.
We propose a novel estimand and its inference framework for dynamic regimes defined during an initial period of intensive longitudinal observation. The framework is implemented using the parametric g-formula with two distinct modelling components: (1) an in-period model for time-varying covariates, treatment decisions, and outcomes during longitudinal observation, and (2) a post-observation outcome model for outcome occurrence conditional on covariates and treatment history at the end of longitudinal covariate observation. These components are connected through sequential simulation, enabling estimation of causal effects beyond the period in which longitudinal covariates are observed.
We applied this framework to an observational study comparing dynamic blood transfusion strategies defined by hemoglobin thresholds in ICU patients, with 28-day mortality as the outcome. The median length of ICU stay was only 4.9 days, substantially shorter than the 28-day follow-up period. Applying the proposed framework, the estimated 28-day mortality was 16.8% under a restrictive strategy and 15.4% under a liberal strategy (RD −1.3%, 95% CI −3.2% to 0.5%). This example illustrates how causal effects of policy-relevant dynamic treatment strategies defined during intensive longitudinal observation can be efficiently evaluated over longer follow-up.